Combination of ultrafast dynamic contrast-enhanced MRI-based radiomics and artificial neural network in assessing BI-RADS 4 breast lesions: Potential to avoid unnecessary biopsies

Abstract

ObjectivesTo investigate whether combining radiomics extracted from ultrafast dynamic contrast-enhanced MRI (DCE-MRI) with an artificial neural network enables differentiation of MR BI-RADS 4 breast lesions and thereby avoids false-positive biopsies.MethodsThis retrospective study consecutively included patients with MR BI-RADS 4 lesions. The ultrafast imaging was performed using Differential sub-sampling with cartesian ordering (DISCO) technique and the tenth and fifteenth postcontrast DISCO images (DISCO-10 and DISCO-15) were selected for further analysis. An experienced radiologist used freely available software (FAE) to perform radiomics extraction. After principal component analysis (PCA), a multilayer perceptron artificial neural network (ANN) to distinguish between malignant and benign lesions was developed and tested using a random allocation approach. ROC analysis was performed to evaluate the diagnostic performance.Results173 patients (mean age 43.1 years, range 18–69 years) with 182 lesions (95 benign, 87 malignant) were included. Three types of independent principal components were obtained from the radiomics based on DISCO-10, DISCO-15, and their combination, respectively. In the testing dataset, ANN models showed excellent diagnostic performance with AUC values of 0.915-0.956. Applying the high-sensitivity cutoffs identified in the training dataset demonstrated the potential to reduce the number of unnecessary biopsies by 63.33%-83.33% at the price of one false-negative diagnosis within the testing dataset.ConclusionsThe ultrafast DCE-MRI radiomics-based machine learning model could classify MR BI-RADS category 4 lesions into benign or malignant, highlighting its potential for future application as a new tool for clinical diagnosis.

Ultrafast Dynamic Contrast‐Enhanced MRI of the Breast: From Theory to Practice

“…Using information from this processing, a UF‐DCE‐based deep learning model achieved an AUC of 0.811, while a model also including information from T2‐weighted imaging, apparent diffusion coefficient mapping, patient age, and BRCA status achieved an AUC of 0.852 65 . Other groups have used deep‐learning for lesion detection, 66 malignancy discrimination, 67 classification of BI‐RADS 4 lesions, 34 and exclusion of normal scans 31 . The tedious work of segmentation and quantification of tumor‐related vessels can potentially be automated using deep‐learning 68 (Fig.…”

“…To improve the temporal resolution of MRI, many research groups have proposed parameters balancing temporal and spatial resolution tradeoffs. 3 Table 1 shows acceleration sequences, temporal resolution, and spatial resolution for reports published after 2020 [6][7][8]10,11,13,14,[18][19][20][21][22][23][24][25][26][27][28][29][30][31][32][33][34][35] (Table 1). Most algorithms fit into one of the following technique categories: view-sharing, conventional gradient-recalled-echo (GRE), or compressed sensing (CS).…”